How We Give a Website Its GEO Score
First, what GEO actually does. A user types a question into ChatGPT or Perplexity and gets a paragraph back. GEO deals with whether you are in that answer, and whether you are described correctly when you are.
It is measured differently from SEO. SEO looks at rankings and clicks: each result has one position, and your rank is unambiguous. A generative engine returns a whole passage. You might be written into it, listed fourth, described as the wrong industry, or left out entirely. So the smallest unit GEO counts is an extractable fact block: a piece of content that can be lifted out on its own and still make sense.
Every threshold in the rubric below maps to a measured figure from published empirical research. We do not give advice based on industry anecdotes. If we cannot cite the basis for an item, it does not go into a client report.
Every score has to trace back to evidence.
- 01
Can machines read it?
We start with crawl conditions and confirm the site's basic entry points.
- 02
Can the content answer?
We check whether the content answers the question and offers clear evidence.
- 03
Does the answer cite it?
We record results under strict sampling rules and keep observations separate from promises.
Three stages, three different problems
Between a user's question and your content appearing in the answer, there are three gates. Each is decided by different factors, and each gives you different levers.
Does the question trigger a web search?
Decided by question type. Time-sensitive and decision-making questions are more likely to trigger one.
Pick the right questions to target. Vague questions trigger retrieval less often.
Which sources make the citation list?
Decided by domain authority, site type and crawlability.
Page technical setup and where you publish. This is your ticket into the candidate pool.
Which content actually makes it into the answer?
Decided by query relevance, length, structure and evidence density.
Content engineering. This is where the gap opens up.
Being cited often and actually being used are two metrics to count separately. In cross-platform experiments, Perplexity cited an average of 16.35 sources per query, Google 12.06, and ChatGPT only 6.88. But the average influence of a single ChatGPT citation was 0.2567, 5.64 times Google's and 4.68 times Perplexity's. Fewer citations, deeper use.
Six dimensions, 100 points
Each threshold comes with the measured evidence behind it. Every recommendation in a client report traces back to a row in this table.
Crawlability
15 ptsIf you can't get into the candidate pool, everything after scores zero. This item checks four things.
- HTTP status is 200
- No noindex
- A canonical tag is present
- Body text is present in the static HTML
The last one is where things most often go wrong. On a purely client-side rendered site, people see the content in a browser, but most AI crawlers get a blank page. It is the first thing we check in an audit.
At the site level, we also check whether robots.txt blocks GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Bytespider, Baiduspider or Google-Extended, and whether sitemap.xml and llms.txt exist.
Content length
15 pts- 1,500+ words: full marks
- 1,000+ words: 85%
- 600+ words: 60%
- 300+ words: 35%
1,000 words is a threshold, not a target. Word count has no value in itself; AI needs enough material to cut out reusable passages.
Structure
20 pts- Exactly one H1
- At least 6 H2s
- At least 25 paragraphs
- List density of at least 0.35
List density is calculated as li divided by the sum of li and p.
Extractable blocks
25 pts · heaviest weight| Block type | Weight | Influence gain |
|---|---|---|
| Definition | 6 | +57.3% |
| Numeric facts | 6 | +61.6% |
| Comparison | 5 | +55.3% |
| How-to steps | 5 | +41.2% |
| FAQ | 3 | No significant gain; helps Q&A recall |
Average influence by page type: definition pages 0.1531, comparison pages 0.1524, pure reference pages only 0.0529.
One result runs against intuition. Pages in a pure Q&A format show an influence of -5.7%. Formatting content as Q&A does not help. What helps is having definitions, numbers, structure and query relevance.
Authority signals
15 pts- A visible publish or update date
- An author
- At least 3 external citations
- JSON-LD structured data
Query relevance
10 pts · strongest predictorChecks whether the title, H1 and H2s use the actual wording of the target question. It is worth only 10 points, yet it is the strongest predictor of the six.
AI is changing the world
What content AI search cites first: a structured analysis of 23,745 citation records
The difference: the second uses words people actually type when they ask.
Breaking down a robots.txt
robots.txt is the shortest file on your site, and the easiest one to get two things wrong in at once. It serves two groups of visitors: traditional search engines and AI crawlers. Below, we use our own setup as the example.
The SEO basics
The first block is the global rule: allow crawling, then block directories that should not be indexed.
User-agent: *
Allow: /
Disallow: /admin/
Disallow: /private/
Disallow: /*.log$
Disallow: /archive/
Disallow: /docs/
Admin pages, private directories, log files, archives and internal docs. If these get indexed, they only dilute how search engines rate your site's overall quality, and they can expose paths that should never be public.
The second block blocks SEO tool crawlers.
User-agent: SemrushBot
Disallow: /
User-agent: AhrefsBot
Disallow: /
User-agent: MJ12bot
Disallow: /
User-agent: DotBot
Disallow: /
User-agent: BLEXBot
Disallow: /
User-agent: DataForSeoBot
Disallow: /
These crawlers bring no traffic. They crawl your site to fill their own backlink databases, which also lets competitors look up your keyword strategy. Blocking them saves server resources and leaks less competitive information.
This is a trade-off. Once you block Ahrefs, using Ahrefs to analyze your own site is affected too. If a client uses these tools for competitor research, we ask first before deciding whether to block them.
The third block allows the major search engines and sets a crawl delay.
User-agent: Googlebot
Allow: /
Crawl-delay: 2
User-agent: Bingbot
Allow: /
Crawl-delay: 3
Sitemap: https://oceanads.org/sitemap.xml
The Sitemap line works anywhere in the file. It tells crawlers where to get the full list of URLs. Without it, crawlers can only follow links level by level, and deep pages are easily missed.
The GEO basics
This part is your ticket into GEO. If robots.txt blocks AI crawlers, you have opted out of the AI answer candidate pool, and no amount of content work will reach it.
At minimum, allow these.
| Crawler | Owner | If you block it |
|---|---|---|
| GPTBot | OpenAI | ChatGPT retrieval can't reach your pages |
| ChatGPT-User | OpenAI | When users paste your URL into a chat, it can't read the content |
| OAI-SearchBot | OpenAI | You are missing from the ChatGPT search index |
| ClaudeBot | Anthropic | Claude retrieval can't reach your pages |
| anthropic-ai | Anthropic | Same as above; legacy user agent |
| PerplexityBot | Perplexity | Perplexity cites the most sources, so missing out costs the most |
| Google-Extended | Affects use in Gemini and AI Overviews | |
| Applebot-Extended | Apple | Affects use in Apple Intelligence |
| CCBot | Common Crawl | Affects open-source corpora, a training source for most models |
| Bytespider | ByteDance | Affects platforms such as Doubao |
| Meta-ExternalAgent | Meta | Affects Meta AI |
The syntax is simple.
# AI search engine crawlers: full access
User-agent: GPTBot
Allow: /
Crawl-delay: 5
User-agent: PerplexityBot
Allow: /
Crawl-delay: 5
User-agent: ClaudeBot
Allow: /
Crawl-delay: 5
A 5-second crawl delay is a compromise. With none, heavy crawling puts load on the server; set it too long and new content reaches AI indexes more slowly. Static sites carry little load, so 5 seconds is enough.
One crawler needs special handling. FacebookBot fetches preview images for share cards. It does not need to read your article text.
User-agent: FacebookBot
Allow: /images/og-
Disallow: /blog/
Disallow: /images/
Allow only images whose names start with og, and block everything else. Share cards still display, and article content stays put.
Two common mistakes
The usual reason is worry that content will be used for training. The concern is valid, but blocking everything is the wrong fix. Once you do, you disappear from ChatGPT and Perplexity answers entirely, with no chance of even being described correctly. A more sensible approach is to allow crawling and spell out usage terms in llms.txt: which content may not be copied in full, which images may not be used for derivatives, and that case studies and pricing may not be used to generate competitor copy.
Allowing crawlers only makes you eligible for the candidate pool. Whether the content they fetch has definitions, numbers and on-topic headings is a matter for the six other dimensions.
Write content as evidence pages
For every question you target, the page needs at least these.
- One definition sentence that can be lifted as is
- Two to five numbers, with units and sources
- One comparison block, ideally a table
- One step-by-step block
- Scenarios where it fits and where it doesn't
Stating clearly who it is not for actually raises credibility.
Sampling discipline
A result without the original answer, a screenshot or a record of the sampling environment cannot be labeled a real sample.
| Grade | Meaning |
|---|---|
| A | Real manual or browser sample, with the original answer text and source links |
| B | Reproducible API sample; must state whether web search was on |
| C | Synthetic demonstration or inference |
| D | Pending review |
Most official APIs do not search the web by default, so they measure brand awareness in the model's parametric knowledge; only the web client performs retrieval. Never combine the two when calculating mention rate.
The same product draws on systematically different sources across clients. In testing, one platform's web and app versions shared only 24.5% of their sources.
One platform's results cannot stand in for what every AI says. Average mention rates across markets explain nothing, so each gets its own table in the report.
Mention rate, first-position rate and Top 3, description accuracy, citation quality. Looking at mention rate alone badly overstates results.
Higher confidence requires a baseline window, an observation window, control question sets, competitor comparisons and a log of external events. When those conditions aren't met, we say so plainly.
We follow platform terms of service: no bulk scraping, no simulated logins. Sampling has a cost, and the metrics themselves are noisy.
The order of priorities
Work tickets are ordered this way, never the reverse.
| Rank | Category | What it covers |
|---|---|---|
| 1 | P0 Entry blockers | Blocked in robots, empty client-rendered shell pages, no sitemap, no structured data, inaccessible pages |
| 2 | P0 Factual errors | AI gets the brand wrong, confuses it with another company, or misstates prices or credentials |
| 3 | P1 Missing extractable blocks | The site lacks definitions, numbers, comparisons or steps |
| 4 | P1 No page to answer the question | High-value questions have no matching page on the site |
| 5 | P1 External source building | Built out on the domains AI actually cites in our samples, not on gut feeling |
| 6 | P2 Long-tail expansion | Extended question sets and content coverage |
External sources rank P1 because of a set of hard numbers. Across 187,818 deduplicated citations, brand websites and corporate sites together made up only 1.37%. Content platforms made up 16.4%, general news media 13.6%, and business recommendation and ranking sites 9.1%.
Your website is the source of facts: it keeps AI's description of you accurate. The citations come from elsewhere. Raising your site from 60 to 90 points is worth less at the margin than getting one entry on a ranking site.
When setting priorities, don't treat SEO rankings, website traffic or brand awareness as the GEO priority order. Having a page on your site and AI being able to cite it accurately are two different things.
What we actually do in one GEO engagement cycle
The question clients ask most is "What do you actually deliver?" This section lays out each cycle's work. Every step has a deliverable and a way to verify it.
Site crawl and page-by-page audit
Deliverable · per-page score sheet
with site-wide gap stats
We crawl the whole site and score every page on all six dimensions. This step surfaces the critical problems first: whether the site is an empty client-rendered shell, whether robots.txt shuts AI crawlers out, whether the Chinese and English page counts match, and which pages lack which block types.
Build the brand fact card
Deliverable · facts card
each item evidence-graded
This document is the input for every asset that follows: official name, common misspellings, product lines, target users, key figures, and where it fits and doesn't. Each item gets an evidence grade.
Fields we can't verify are marked "to be confirmed" rather than filled in with assumptions. Founding date, registered legal entity and client cases that can be named are the fields most often left blank, and the client needs to fill them in.
One sentence in the fact card matters most: the brand definition. It must be identical, word for word, in four places: the homepage hero, the About page, the JSON-LD description, and llms.txt. Inconsistent wording is the number one reason AI descriptions of a brand drift.
Build the question bank
Deliverable · seven question groups
each tagged by market
The question bank decides what we sample, what content we write and how metrics are calculated. Questions fall into seven groups: recommendation, comparison, alternatives, pricing, risk, brand verification, and use case.
Each question is tagged by market. Chinese questions won't be run on Perplexity, and English questions won't be run on Doubao. For overseas markets, questions must be phrased the way native English speakers ask, not machine-translated from Chinese.
Brand verification questions name the brand, so the answer will always repeat it. These go into a separate brand awareness bucket and are kept out of visibility metrics.
AI answer sampling
Deliverable · raw sample files
with answer text and environment
Platforms with a public API are sampled through the API; overseas coverage includes Gemini, ChatGPT, Claude, Grok and Perplexity. Platforms without a public web-search API are sampled manually or in a browser: we export a sampling sheet, fill it in, then feed it back in.
Each record captures the original answer, whether web search was on, the client type and the sampling time.
Generate work tickets
Deliverable · ticket sheet
HTML and CSV
Diagnostic results become work tickets, the single source of truth for execution status. Each ticket carries five things: the measured figure it traces back to, the owner role, the effort, the time window, and the acceptance criteria.
Tickets fall into seven work packages: entity disambiguation, page technical, content planning, heading system, knowledge base, external evidence, and performance monitoring.
Produce ready-to-deploy assets
Deliverable · assets folder
Chinese and English separate
| Asset | Purpose |
|---|---|
| llms.txt | Official fact index, placed directly in the site root |
| JSON-LD | Structured data such as Organization, FAQPage, Article and BreadcrumbList, pasted into the head |
| Definition block | Placed below the hero tagline; the tagline stays, so conversions are unaffected |
| FAQ block | Answers must be visible in the static HTML, not hidden in JS-only accordions |
| Content outline | One per target question, with title options, a section skeleton, word count and block requirements |
Structural assets are generated deterministically by code, so no schema fields get missed. Article body text is written by people from the outline.
Automated verification
Deliverable · verification sheet
ticket status written back
We re-crawl the site, run the checks, and write ticket status back. Finished items flip to done; regressions flip back to to-do.
This step is the line between a service and advice. Anything that can be verified automatically doesn't rely on someone saying it's done. Items code can't judge, such as whether a wiki entry passed review, are marked for manual confirmation.
Delivery package
Deliverable · one package per cycle
ready to archive
Each cycle is packaged with an overview, diagnostic report, action plan, ticket sheet, verification sheet, assets folder and delivery notes. The ticket sheet is also exported as CSV, so you can import it straight into a project management tool.
We recommend running page audits weekly and answer sampling every two weeks or monthly. Sampling has a cost and the metrics are noisy; run it too often and you won't see a signal.
What we don't promise
No one can guarantee that a given platform will cite a given page. GEO improves discoverability, verifiability and extractability. What it raises is the probability.
Brand facts we can't verify are always marked "to be confirmed", never filled in with assumptions. Fields like founding date, registered legal entity and client cases that can be named stay blank until the client supplies them.
We don't do bulk scraping that circumvents platform limits, simulated logins or traffic inflation.
To see this rubric in practice, read our full audit report on our own website, including sampling results showing we barely appear in AI answers.
FAQ
How many points is each of the six GEO scoring dimensions worth?
100 points in total. Crawlability 15, content length 15, structure 20, extractable blocks 25, authority signals 15, query relevance 10. Extractable blocks carry the most weight. Query relevance is worth only 10 points, yet it is the strongest single predictor of influence, with a correlation of 0.432.
What kind of content is most likely to be absorbed into AI answers?
Ranked by measured influence gain: numeric facts +61.6%, definitions +57.3%, comparisons +55.3%, how-to steps +41.2%. Pages in a pure Q&A format show -5.7%. Formatting content as Q&A does not help on its own.
Can API sampling and web-app sampling be combined?
No. Most official APIs do not search the web by default, so they measure brand awareness in the model's parametric knowledge; only the web client performs retrieval. Mention rates for the two must be recorded separately. The web version and the mobile app must also be kept apart: in testing, different clients of the same platform shared only 24.5% of their sources.
Why is building external sources ranked P1?
Across 187,818 deduplicated citations, brand websites and corporate sites together made up only 1.37%. Content platforms made up 16.4%, general news media 13.6%, and business recommendation and ranking sites 9.1%. Your website is the source of facts; the citations come from elsewhere. Raising your site from 60 to 90 points is worth less at the margin than getting one entry on a ranking site.
Run this rubric on your website
Send us your URL. We'll turn the six-dimension scores, extractable block gaps and sampling results into a report you can actually read. Every recommendation states its basis.